The Agent Skills Directory

[NO_CODE]: The skill consists solely of markdown instructions (README.md, SKILL.md) and reference documentation (references/). It does not include any scripts, executable binaries, or configuration files that trigger side effects.
[SAFE]: The instructions are designed to improve safety and correctness by enforcing 'context walls' between agents and requiring 'evidence-anchoring' (verbatim quotes) for any quality claims. There are no patterns suggesting data exfiltration, credential theft, or unauthorized command execution.
[PROMPT_INJECTION]: While the skill contains strong instructions (e.g., 'MUST oppose', 'Invariants (low freedom)'), these are standard operational constraints for the multi-agent pattern and do not attempt to bypass the underlying LLM safety guidelines or override the system prompt in a malicious way.
[DATA_EXFILTRATION]: No network tools or external data transfer mechanisms are used or requested. The skill operates within the existing environment provided by the platform.
[COMMAND_EXECUTION]: The skill mentions running verification commands (e.g., npm test), but it does not provide arbitrary command execution logic. It expects the user or orchestrator to define these based on the specific artifact being tested.

tribunal